论文阅读：Non-local Neural Networks

论文地址：https://arxiv.org/pdf/1711.07971.pdf

内容简介

主要讨论下提出来的non-local block

non-local block

直译下来叫“非局部块”。长这样：

既然叫“非局部”那就看看什么叫做“局部”。所谓局部指的就是感受野，比如卷积核的大小。一般来说我们使用的都是如3×3的小尺寸卷积核，所以只能提取小范围的信息。

而为了克服这种局面，获取更大范围(比如两个较远像素)之间蕴含的信息，就可以用文章中提出的non local结构。

具体的流程是这样的：

对于输入的特征图(假设是1024通道)，用1个1×1卷积将其降维至512通道(更通用的讲，降一半)，以供备用，记为ggg
同样，分别用2个1×1卷积处理输入特征图，记为θ\thetaθ和ϕ\phiϕ
将θ\thetaθ和ϕ\phiϕ做矩阵乘，得到的结果类似于协方差矩阵，包含自相关性信息
对该结果做softmax，获得的权重就相当于self attention系数
与g相乘，升维，再与原始输入x残差相加，得到最后结果

代码长这样：

class NonLocal(nn.Module):def __init__(self, channel):super(NonLocalBlock, self).__init__()self.inter_channel = channel // 2self.conv_phi = nn.Conv2d(channel, self.inter_channel, 1, 1, 0, False)self.conv_theta = nn.Conv2d(channel, self.inter_channel, 1, 1, 0, False)self.conv_g = nn.Conv2d(channel, self.inter_channel, 1, 1, 0, False)self.softmax = nn.Softmax(dim=1)self.conv_mask = nn.Conv2d(self.inter_channel, channel, 1, 1, 0, False)def forward(self, x):b, c, h, w = x.size()x_phi = self.conv_phi(x).view(b, c, -1)x_theta = self.conv_theta(x).view(b, c, -1).permute(0, 2, 1).contiguous()x_g = self.conv_g(x).view(b, c, -1).permute(0, 2, 1).contiguous()mul_theta_phi = torch.matmul(x_theta, x_phi)mul_theta_phi = self.softmax(mul_theta_phi)mul_theta_phi_g = torch.matmul(mul_theta_phi, x_g)mul_theta_phi_g = mul_theta_phi_g.permute(0, 2, 1).contiguous().view(b, self.inter_channel, h, w)mask = self.conv_mask(mul_theta_phi_g)out = mask + x return out

参考

https://mp.weixin.qq.com/s/FFEKWFgScdBZ8snZQLdfFg

论文阅读：Non-local Neural Networks相关推荐

论文阅读2018-Deep Convolutional Neural Networks for breast cancer screening 重点：利用迁移学习三个网络常规化进行分类
论文阅读2018-Deep Convolutional Neural Networks for breast cancer screening 摘要:我们探讨了迁移学习的重要性,并通过实验确定了在训练 ...
[论文阅读笔记]Deep Neural Networks are Easily Fooled:High Confidence Predictions for Unrecognizable Images
Deep Neural Networks are Easily Fooled:High Confidence Predictions for Unrecognizable Images(CVPR201 ...
【论文阅读】Deep Neural Networks for Learning Graph Representations | day14，15
<Deep Neural Networks for Learning Graph Representations>- (AAAI-16)-2016 文章目录一.模型 1.1解决了两个问题 ...
论文阅读：Recurrent Neural Networks for Time Series Forecasting Current Status and Future Directions
typora-copy-images-to: ./ Recurrent Neural Networks for Time Series Forecasting: Current Status and ...
论文阅读：Convolutional Neural Networks for Sentence Classification 卷积神经网络的句子分类
Convolutional Neural Networks for Sentence Classification 卷积神经网络的句子分类目录 Convolutional Neural Networ ...
[ICCV2019论文阅读]Asymmetric Non-local Neural Networks for Semantic Segmentation
作者:Zhen Zhu, Mengde Xu, Song Bai, Tengteng Huang, Xiang Bai 机构:Huazhong University of Science and Te ...
【论文阅读】Siamese Neural Networks for One-shot Image Recognition
关键词: one-short learning : 待解决的问题只有少量的标注数据,先验知识很匮乏,迁移学习就属于one-short learning的一种 zero-short learning: ...
【论文阅读】Diffusion-Convolutional Neural Networks (DCNNs) | day9、10
30th Conference on Neural Information Processing Systems (NIPS 2016) 文章目录一.介绍优势: 二.模型三.缺点用于图结构 ...
论文阅读（Improving neural networks by preventing co-adaptation of feature detectors ）
显然看不懂英文然后出于各种原因最终被迫搜索了翻译版本,如有错误欢迎指出介绍主体: DropOut 目的: 对战过拟合的基础上提升速度和学习能力(力速双A✔) 思路: 通过在每个训练案例上随机省略一般 ...

论文阅读：Non-local Neural Networks

内容简介

non-local block

参考

论文阅读：Non-local Neural Networks相关推荐

最新文章

热门文章